Evaluating the impact of Radiology Reports Structure on AI-Powered Radiology Report Generation Systems
by Zaheer Babar
Radiology reports play an essential role in diagnosing and monitoring various diseases and conditions, from pneumonia to lung cancer and bone conditions. The ability to convey findings clearly and comprehensively is paramount, and producing well-structured, clear, and clinically well-focused radiology reports is essential for high-quality patient diagnosis and care. High-quality patient diagnosis and care can be achieved using a computer-aided radiology report system, which assists radiologists in producing well-structured, clear, and clinically well-focused radiology reports. Deep learning has made significant strides in image caption generation, but it has remained a highly challenging task in the medical domain.
One main challenge is understanding and linking complicated medical observations detected in given images with accurate natural language descriptions. Radiologists follow a standard way of writing these reports, describing a fixed set of diseases and conditions, indicating whether it is normal or abnormal. As a result, medical reports usually overlap with each other due to the common content of anatomy. This standardized way of reporting makes it challenging for the machine learning model to capture the prominent problems and abnormalities indicated in radiology reports. This impact can be felt across various aspects of the task, ranging from the utilization of validation metrics to the performance of the model and the use of different components within it. In this thesis, we study this impact on different levels and demonstrate that our research will lead to reliable progress in automatic radiology report generation.